Export alignment spectrum provenance links

[htsugawa]

Adds two Console text exports that make an alignment spot traceable back to the per-file peak it came from: AlignResult-*.mdpeakid.tsv, a compact matrix from each MasterAlignmentID and sample to the source MasterPeakID, and the optional AlignResult-*.mdprovenance.tsv audit sidecar behind a new Detailed alignment provenance config key. Without them there is no way to say which .mdpeak row a given aligned value came from, which is the link the companion catalog needs in order to attach an annotation claim to the evidence behind it.

This branch has been sitting unmerged since 2026-09-01 and had fallen ten commits behind, so it is rebased onto current master (f36aea416).

The rebase

One textual conflict, in ConfigParser.cs, where both sides added a new reader method at the same position: master added ReadLbmAnnotatorPriority (#783), the branch adds ReadDetailedAlignmentProvenance. Both are needed, so both are kept side by side, each verbatim as its own side wrote it. LcmsProcess.cs auto-merged because the two sides touch different regions — master the annotator setup, the branch the export block and the ExecuteAsync signature.

ConfigParser.cs uses method.ToLower() nine times and ToLowerInvariant() once, the once being master's new method. Each side is left as written: a rebase should not smuggle in a normalization neither side authored. The nine ToLower() sites are a real latent bug — under a Turkish or Azeri culture 'I'.ToLower() gives dotless ı, so a config key silently stops matching and the run proceeds with the default — but they belong in their own change.

The rebase changes no existing output

Measured, not reasoned. Same demo, same method file, this branch against plain master:

Output	Result
7 per-file .mdpeak, 7 per-file .mdmsp	byte-identical
.mdalign, alignment .mdmsp, .qa.tsv	byte-identical
.mzTab	differs only in the MTD mzTab-ID line, which is the run timestamp
Extra files	.mdpeakid.tsv and .mdprovenance.tsv, nothing else

The alignment spot count on this demo moved from 2512 to 2553 relative to the older checked-in reference output. That is master's, not this branch's: #773 made sample max / blank average also set FoldChangeForBlankFiltering, which changes blank filtering, which runs at the alignment stage. Peak-detection columns are unchanged in every row — Peak ID, RT (min), Precursor m/z, Height, Area, S/N, Scan all show zero differing rows — and the .mdpeak differences are confined to the annotation block, from #783, #784, #785 and #789.

Three things the sidecar asserted without measuring

Found by review of the rebase and fixed in 7a7b8c28d. All three were invisible to the original tests, because both asserted substrings of a row that had a real source peak and stopped before the mz column.

mz was backwards in every row. The in-memory overload read ChromXsTop.Mz, the light overload read Mass. On an aligned peak ChromXsTop carries the chromatogram axis and the gap filler resets it, so the column reported a sentinel for every member that did have a source peak and a real value only for the gap-filled ones. On the FastLC demo: 7879 of 7879 has_source_peak=true rows carried mz=-1, and 9992 of 9992 gap-filled rows carried a plausible m/z. Both overloads now read Mass, which is what the .mdalign MZ column uses.

Gap-filled rows published a false raw-scan pointer. The gap filler resets the peak ids but leaves the spectrum ids at their default 0, and 0 is a real scan index. Every gap-filled row therefore claimed ms1_raw_spectrum_id=0, ms1_raw_spectrum_id_top=0 and ms2_raw_spectrum_id=0. In a file whose purpose is provenance that points an auditor at a spectrum belonging to a different peak. All five raw-spectrum columns are now withheld when there is no source peak, in both overloads. The light overload had additionally gated ms2_raw_spectrum_ids on MS2RawSpectrumID >= 0, which is true for that same 0.

The two files disagreed about the same cell. The sidecar wrote -2 for a gap-filled peak id where the matrix wrote -1. They now agree on -1, and the distinction is preserved by name in a new peak_origin column valued detected, gap_filled or absent. Same lesson as #785: name the state instead of encoding it in a magic number.

One behavioural change to existing output, in alignment-light mode only

The QA-matrix block's six accessors move from an unconditional new LegacyQuantValueAccessor(...) to a factory that returns AlignmentLightQuantValueAccessor when the light store is in use. This is a bug fix rather than part of the export: in light mode each spot holds only its representative peak, so LegacyQuantValueAccessor, which iterates AlignedPeakProperties, emitted one column's worth of values for the whole matrix. Flagging it explicitly because the commit subject does not suggest the QA matrix would change.

That switch also exposed a divergence in the light-mode Reference matched column, fixed here: it derived from RepresentativeLibraryID >= 0, but a suggested annotation carries a library id too, so a precursor-only or low-score suggestion counted as a reference match in light mode while normal mode reported false through AnnotationName.IsReferenceMatched, the predicate #784 centralised. The store now persists the representative's own IsReferenceMatched verdict. RecordSize goes 102 to 103; the file is a per-run temporary spill from Path.GetTempFileName, so the layout carries no cross-version obligation.

Verification

• Build: Release, net48, 0 errors.
• MsdialCoreTests: 316/316 (311 before, plus five new covering gap-filled and unwritten-light rows, the mz-from-Mass rule, and the reference-matched verdict).
• MsdialCoreTestAppTests: 13/13, including a new ReadDetailedAlignmentProvenance test next to master's ReadLbmAnnotatorPriority test. That is the project that actually compiles the rebased ConfigParser, and running it was the one piece of evidence the original green result did not cover.
• End to end on the FastLC demo, ingested into the companion catalog: every member with a source peak carries a usable m/z, no member without one carries a scan index, and peak_origin splits 17871 members into 7879 detected and 9992 gap_filled.

Two guards so this class of defect cannot return quietly: WriteMember now goes through a WriteRow that refuses a field count the header does not declare, since the header and the two value lists are three hand-maintained arrays; and the tests assert whole lines rather than substrings.

Deliberately not in this PR

• Gating .mdpeakid.tsv behind a config flag. It is currently unconditional, and in light mode it adds an O(spots x files) seek-per-cell pass — which is exactly the regime light mode exists for. Worth a follow-up; the matrix is the authoritative join key, so defaulting it off would need thought.
• Converting the nine method.ToLower() sites to the invariant form.
• Splitting the QA-accessor fix into its own commit.

🤖 Generated with Claude Code